[Mlir-commits] [mlir] [mlir][scf] Do not read non-constant loop bounds when unrolling (PR #217392)

[Alessandro Potenza]

alepot55 wrote:

The `amdgpu-clang-flang` failure above is not from this patch, and the builder has already recovered on its own: [build 2924](https://lab.llvm.org/buildbot/#/builders/234/builds/2924) is green with this commit still in the tree and nothing reverted.

Three things say the same:

- the failing test is `libomptarget :: amdgpu-amd-amdhsa :: tools/omp-kernel-replay/record-replay-diff-teams-threads.cpp`, an OpenMP offload runtime test, while this commit touches only `mlir/lib/Dialect/SCF/Utils/Utils.cpp` and `mlir/test/Dialect/SCF/loop-unroll.mlir`;
- clang's OpenMP offload path does not go through MLIR, so there is no route from SCF loop unrolling to that runtime;
- the same step already failed on this builder at [build 2919](https://lab.llvm.org/buildbot/#/builders/234/builds/2919), before this commit existed, with 2920 to 2922 green in between.

Flagging it only because the blamelist for 2923 was this commit alone, so anyone reading the notification would land here first.

https://github.com/llvm/llvm-project/pull/217392